THE economic forces that are driving the cellular industry

Transcription

1 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. XX, MAY Balancing Specral Efficiency, Energy Consumpion, and Fairness in Fuure Heerogeneous Wireless Sysems wih Reconfigurable Devices Rahul Amin, Suden Member, IEEE, Jim Marin, Member, IEEE, Juan Deaon, Member, IEEE, Luiz A. DaSilva, Senior Member, IEEE, Amr Hussien, Suden Member, IEEE, and Ahmed Elawil, Member, IEEE Absrac In his paper, we presen an approach o managing resources in a large-scale heerogeneous wireless nework ha suppors reconfigurable devices. The sysem under sudy embodies inerneworking conceps requiring independen wireless neworks o cooperae in order o provide a unified nework o users. We propose a muli-aribue scheduling algorihm implemened by a cenral Global Resource Conroller (GRC) ha manages he resources of several differen auonomous wireless sysems. The aribues considered by he muli-aribue opimizaion funcion consis of sysem specral efficiency, baery lifeime of each user (or overall energy consumpion), and insananeous and long-erm fairness for each user in he sysem. To compue he relaive imporance of each aribue, we use he Analyical Hierarchy Process (AHP) ha akes inerview responses from wireless nework providers as inpu and generaes weigh assignmens for each aribue in our opimizaion problem. Through Malab/CPLEX based simulaions, we show an increase in a muli-aribue sysem uiliy measure of up o 57% for our algorihm compared o oher widely sudied resource allocaion algorihms including Max-Sum Rae, Proporional Fair, Max-Min Fair and Min Power. Index Terms heerogeneous wireless neworks, reconfigurable radios, scheduling, resource allocaion, nework efficiency. I. INTRODUCTION THE economic forces ha are driving he cellular indusry are reducing he number of cellular providers bu causing heir wireless neworks o become large, heerogeneous sysems based on numerous wireless echnologies a various lifecycle sages. Unil recenly, echnology was he primary impedimen o achieving universal, broadband wireless services ha involve muliple radio access echnologies (RATs). Today, arguably one of he mos significan impedimens o achieving highly efficien unified wireless sysems are he afer-effecs of Manuscrip received April 10, 2012; revised November 2, Rahul Amin is wih he Deparmen of Elecrical and Compuer Engineering, Clemson Universiy, Clemson, SC, 29634, USA ( ramin@clemson.edu). Jim Marin is wih he School of Compuing, Clemson Universiy, Clemson, SC, 29634, USA ( jmary@clemson.edu). Juan Deaon is wih Idaho Naional Lab, Idaho Falls, ID, 83415, USA and a recen Ph.D. graduae from Deparmen of Elecrical and Compuer Engineering, Virginia Tech, Blacksburg, VA, ( juan.deaon@gmail.com). Luiz A. DaSilva is wih he Deparmen of Elecrical and Compuer Engineering, Virginia Tech, Blacksburg, VA, ( ldasilva@v.edu). He is also wih CVTR, Triniy College Dublin. Amr Hussien and Ahmed Elawil are wih he Deparmen of Elecrical and Compuer Engineering, Universiy of California, Irvine, CA, ( ahussien@uci.edu, aelawil@uci.edu) /00$00.00 c 2013 IEEE aniquaed specrum allocaion regulaory policies. The effec is ha in many geographic areas licensed specrum is likely o be underuilized even hough demand is expeced o exceed curren specrum capaciy as early as 2014 [1, 2]. This has sparked renewed ineres in echniques o improve specral efficiency, including cogniive radios and neworks ha can adap heir behaviors o make efficien use of open or unused specrum. A he same ime, environmenal concerns and user device requiremens have elevaed he imporance of energy efficien neworks and devices. As wireless operaors have learned, a handheld device s baery efficiency is a very visible aribue of an operaor s services [3]. Unforunaely, in many siuaions, mehods for improving specral efficiency direcly lead o an increase in energy consumpion. A hird conflicing dimension o he resource allocaion problem is ensuring fair allocaion across users. The wireless sysem design and he subsequen performance modeling and evaluaion ha are presened in his paper direcly address his challenging dilemma hrough he use of join opimizaion. A large amoun of prior work has explored a number of issues surrounding he delicae relaionship beween ransmi power and subsequen effecs on compeing devices [4 6]. Recen work involving heerogeneous cellular sysems based on pico-cell and femo-cell deploymens also consider he power issue o miigae iner and inra channel inerference [7 9]. Prior join opimizaion-based resource allocaion approaches for wireless neworks consider rade-offs beween specral efficiency and fairness objecives subjec o maximum ransmi power consrains [10 12]. Recenly, due o a renewed ineres in green communicaions and end-users increased expecaions from mobile device baery life, researchers have sared focusing on minimizing overall energy consumpion subjec o fairness consrains and oher nework-efficiency requiremens, such as hroughpu and delay [13 15]. However, o he bes of our knowledge, none of hese works looked a he rade-offs surrounding power, specral efficiency, and fairness in large-scale heerogeneous wireless neworks ha involve reconfigurable end-user devices. Curren cellular sysems deploy a hierarchy of resource conrollers. Each device, along wih is assigned base saion (BS), independenly ries o opimize he resource allocaion process wihin is own domain, generally ignoring impacs of colocaed heerogeneous wireless neworks. Localized resource allocaion decisions will usually no lead o opimal resource

2 2 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. XX, MAY 2013 usage. In fac, [16] shows ha he selfish approach can resul in non Pareo-opimal bandwidh allocaion as compared o he case where a cenralized eniy performs nework-wide resource allocaion. Significan improvemens in efficiency resul when he resource managemen process joinly considers he disribuion of resources across nework echnologies, reaping he benefis of muli-access nework diversiy. A he nework level, several archiecures and frameworks o suppor hybrid or heerogeneous neworks have been suggesed, which include IEEE , IEEE P1900.4, 3GPP s Common Radio Resource Managemen, Join Radio Resource Managemen and Muli-access Radio Resource Managemen, and IETF s flow mobiliy sandards [17 22]. A common aribue of all hese frameworks is ha he local resource managers of differen RATs inerac wih a cenralized eniy o joinly opimize he resource allocaion process. In our prior work involving a heerogeneous wireless sysem based on a global (i.e. cenralized) resource conroller, we sudied he specral efficiency and energy consumpion radeoffs as he reconfiguraion capabiliies of devices were varied [23]. Resuls from a paricular scenario of his prior work suggess ha i is possible o increase specral efficiency by up o 75% bu a he cos of wice he energy requiremen of devices. In he research presened in his paper, we exend our prior work o include a join opimizaion-based resource allocaion process ha provides an operaor wih a conrol knob o allow he conflicing demands of specral efficiency, energy consumpion, and fairness o be ailored o mee specific performance goals or policies. We use uiliy heory, and in paricular a weighed sum muli-aribue opimizaion echnique, o se up our nework opimizaion problem [24]. The operaor can hen use economic incenives o align user perceived uiliy wih an operaor s financial uiliy, which is a separae problem and no addressed in he work presened in his paper. We inroduce a muli-aribue resource allocaion algorihm ha accouns for cos associaed wih he use of reconfigurable devices. Based on he well-known uiliy-based opimizaion echnique [25], we use he weighed sum mehod in our algorihm, which maximizes uiliy funcions relaed o muliple nework-efficiency aribues. To provide a concree illusraion of he proposed algorihm, we assume he nework provides a bes-effor daa service wih exensions o suppor real-ime raffic. The exensions include wo componens: 1) admission conrol; and 2) minimum insananeous daa rae per scheduling inerval. The service definiion maches he needs of laency-sensiive applicaions such as over-he-op VoIP and video conferencing ha periodically require small amouns of daa o be delivered by he nework wih igh delay bounds. The main conribuion of he research is an approach for managing resources in large-scale heerogeneous wireless neworks ha involve reconfigurable end-user devices. Moreover, he proposed resource allocaion procedure provides a mechanism ha allows an operaor o configure he relaive imporance given o specral efficiency, energy consumpion, and fairness objecives in a manner ha leads o predicable service levels. The paper is organized as follows. Secion II presens relevan background and moivaions for our work. We describe he sysem model, provide a deailed problem formulaion and discuss he research mehodology in Secion III. We presen and discuss he resuls in Secion IV. Secion V provides conclusions and idenifies possible fuure work. II. BACKGROUND AND MOTIVATIONS Nework selecion algorihms for opimal service delivery over user devices capable of connecing wih several RATs can be caegorized ino several sraegies: (i) decision funcionbased sraegies; (ii) user-cenric sraegies; (iii) muliple aribue decision making sraegies; and (iv) fuzzy logic and neural neworks-based sraegies. All hese sraegies use a se of aribues in he decision making process which are eiher relaed o he user or o he service provider. Some of he user-relaed aribues include achieved hroughpu by each individual user, baery lifeime of each mobile erminal, and QoS parameers such as packe delay, jier and loss. Service provider-relaed aribues include load-balancing, hroughpu fairness amongs users, incurred cos per ransmied daa bye, and overall revenue. The decision funcion-based sraegies use a weighed uiliy funcion ha incorporaes boh user-relaed and service provider-relaed nework selecion aribues [26, 27]. The user-cenric sraegies focus on one or more needs of he user o decide on he choice of curren access nework [28]. Muliple Aribue Decision Making (MADM) deals wih he problem of choosing from a se of alernaives ha are characerized in erms of heir aribues. The mos popular classical MADM mehods are Simple Addiive Weighing (SAW), Technique for Order Preference by Similariy o Ideal Soluion (TOPSIS), and Grey Relaional Analysis (GRA). A comparison of hese models was esablished in [29] wih bandwidh, delay, jier, and BER aribues. I showed ha SAW and TOPSIS provide similar performance under all raffic classes examined. GRA provides slighly higher bandwidh and lower delay o ineracive and background raffic classes. Fuzzy Logic (FL) and Neural Neworks (NN) conceps are applied o choose when and o which nework o hand-off among differen available access neworks when a decision problem conains aribues wih imprecise informaion [30, 31]. All he sraegies for nework selecion algorihms described in he previous paragraph make use of muliple user-relaed or nework provider-relaed aribues. The mehod o deermine he relaive imporance of each aribue under consideraion has significan impac on he soluion space and he implemenaion complexiy of he algorihm. Several papers have looked a muliple weigh combinaions of he aribues based on imprecise end-user preferences [27, 32, 33]. Oher papers have seleced he aribue weighs based on simulaion resuls by deermining he difference in magniude of each aribue and hen assigning each aribue equal imporance [26, 31]. We reduce he soluion space of our algorihm by using responses from nework provider inerviews and he Analyical Hierarchy Process (AHP) [34] o deermine he relaive weighs of each aribue in our opimizaion funcion. Game heory has also been employed o model he nework selecion problem. The auhors of [35] propose a nework selecion scheme o accommodae curren demand and minimize

3 Signal Power Eherne Eh Wireless e s Signal Power Eherne Wireless Signal Power Eherne Wireless AMIN e al.: BALANCING SPECTRAL EFFICIENCY, ENERGY CONSUMPTION, AND FAIRNESS 3 VoIP Media GRC cue IP-CAN 1 IP-CAN 2 GRC IMS Subsysem Sense/Selec Local RANs IP Nework Access GRC Regisraion GSM UMTS LTE HSPA EVDO CDMA Nework Sensing Repor IP-CAN Assignmen RR RR RR RR RR RR IP Nework Access Fig. 1. Sysem Model Fig. 2. Resource Allocaion Procedure handoff while meeing QoS requiremens in a heerogeneous wireless nework, comparing he proposed scheme o TOPSIS. The model in [36] consiss of a game beween access neworks in a converged 4G environmen, o decide which service requess should be accommodaed by each access nework. In [37], he auhors sudy he dynamics of nework selecion in a heerogeneous wireless environmen using evoluionary games. Game heory formulaions model decision-making by auonomous independen agens, while in his paper we focus on a cenral global resource conroller. Sill, he exensive lieraure on game heory for elecommunicaions (e.g., [38]) provides rich ideas on nework and user uiliy when considering muliple aribues for opimizaion. A. Sysem Model III. SYSTEM DESCRIPTION Wih he major nework operaors migraing o IP Muli- Media Service (IMS) core neworks [39], we consider an archiecure based on he 3GPP IMS archiecure [40] shown in Fig. 1. In our archiecural model, we consider cogniive User Equipmen (cue) as an end-user device wih reconfigurable and cogniive capabiliies, which is able o access muliple IP Conneciviy Access Neworks (IP-CANs) individually or simulaneously. Resources are managed hrough he Global Resource Conroller (GRC) according o he objecives of he nework operaor. Through hese objecives, he GRC calculaes cue-ip-can mappings and he rae assignmen per mapping. The base IMS framework naurally exends o suppor muli-radio, adapive devices. From an operaional perspecive, he cue firs mus sense IP-CANs and regiser wih he GRC before ransmiing any daa. We show he procedural flow example of his process in Fig. 2. Firs, he cue senses and scans for available neworks and heir uilizaion. Selecing one of he available IP-CANs, he cue obains an IP Nework connecion, which i uses o communicae wih exernal hoss. We assume ha each end-user ries o use he mos efficien IP-CAN available iniially and follows he following preference order: Wi-Fi, 4G (LTE/WiMAX), 3G (HSPA/EVDO). If he user canno esablish a connecion o his/her firs preference due o reasons such as very high nework load or inerference, hen he/she ries o connec o his/her second preference and his procedure coninues unil he cue can esablish an iniial IP Nework connecion. Nex, he cue discovers, regisers, and communicaes wih he GRC applicaion server, which we assume uses sandard discovery and regisraion procedures as described in [40] 1. Afer a connecion wih he GRC is esablished, he cue delivers periodic sensing informaion of available neworks o he GRC. Upon receiving his periodic sensing informaion from he cue, he GRC is able o calculae he cue-ip-can mappings and he rae assignmen per mapping. This informaion is hen relayed o each cue, which uses i o une is Reconfigurable Radios (RRs) o he appropriae IP- CANs. Afer each RR is configured according o he cue-grc mapping, radio links are esablished wih he associaed IP- CANs for daa ransmission. A picorial represenaion of he ransmission plane is shown in Fig. 3. From he perspecive of he cue applicaions, one TCP/IP sack is managed and scheduled over one or more radio links. The Radio Link Aggregaion funcion is used for packe resequencing and reordering daa from each of he RRs. Each RR manages is own radio link and associaed proocol wih he IP-CANs, which provides a unique IP anchor hrough he corresponding IP-CAN s core nework. In addiion, each RR uses he rae assignmen per cue-ip-can mapping informaion calculaed by he GRC in is resource demand requess o he corresponding IP-CAN s BS. The IP-CAN BSs use he resource demand requess from each cue as guidance in coming up wih heir own local scheduling decisions. Noe ha he GRC makes periodic decisions on large ime scales (seconds or minues), while he BSs of each IP-CAN make scheduling decisions on small ime scales (milliseconds) o accoun for shor-erm flucuaions in conneciviy condiions. While some of he seings are cusomizable for LTE and WiMAX, generally hese IP-CANs generae a schedule every 10 milliseconds. HSPA ypically generaes a schedule every 2 milliseconds and EVDO generaes a schedule every milliseconds. Wi-Fi ypically assigns a channel o he user for 0.5 milliseconds o send one daa frame (which includes he DIFS, Daa, SIFS, ACK mechanism). The inen of he proposed soluion is o le independen schedulers of each IP-CAN work a heir defined schedule inerval duraions. The GRC performs global-level opimizaion (re-associaions) and has o operae a larger 1 Regisraion wih an IMS applicaion server involves a combinaion of DNS lookups wih Diameer auhenicaion procedures (RFC 3588), and SIP signaling.

4 Signal Power Eherne Wireless 4 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. XX, MAY 2013 cue IP-CAN 1 IP-CAN 2 IP-CAN 3 Core 1 Fig. 3. Daa Transmission Plane TCP-IP Radio Link Aggregaion RR-1 RR-2 RR-3 Core 2 Core 3 Inerne Exernal Hoss/Servers ime scales o accoun for issues such as overhead/resul propagaion delay. So, o minimize acual overhead and o make sure ha he cues and BSs of various IP-CANs can use he decisions made by he GRC, a scheduling inerval of 1 second is proposed for he GRC and is used hroughou he paper. B. Opimizaion Aribue Uiliy Funcions The GRC uses a muli-aribue resource allocaion algorihm o deermine he cue-ip-can mappings and he rae assignmen per mapping for each scheduling ime sep. The aribues considered in his algorihm are sysem specral efficiency, boh insananeous and long-erm fairness in erms of daa rae allocaed o each user in he sysem, and baery lifeime of each user in he sysem. We describe he uiliy funcion for each aribue nex using sysem parameers presened in Table I. The noaion exends he noaion presened in [16]. 1) Specral Efficiency: The achievable sysem specral efficiency for ime inerval [, +1], denoed γ, is represened by (1) as he raio of he aggregae rae allocaed o each user in he sysem a ime o he oal specrum used. The rae allocaed o user u U a ime, denoed r u, is represened by (2) and depends on wo parameers: (i) x ua - he cue-ip- CAN assignmen variable a ime, (ii) r ua - he rae allocaed o user u U by BS/AP a A a ime. For each ime sep, he muli-aribue resource allocaion algorihm implemened by he GRC compues boh hese parameers. Noe ha r ua is a funcion of he resource blocks assigned o user u U by BS/AP a A a ime and he suppored modulaion and coding scheme (MCS). A resource block is a minimal resource allocaion uni. Differen IP-CANs use differen erminology when defining a minimal resource allocaion uni (for example, Wi-Fi les users compee for he wireless medium using he CSMA/CA mechanism and les he conenion winner hold he wireless medium for he ime necessary o ransmi a daa frame and ACKs plus any opional conrol frames associaed wih virual carrier sensing, OFMDA-based LTE and WiMAX group welve consecuive subcarriers in he frequency domain Parameer A U BU x ua r ua r ua,max r ua,norm r u γ κ η ua T ω ua ω u m u Descripion TABLE I SYSTEM PARAMETERS Se of BSs/APs for all IP-CANs Se of Users Se of users ha are blocked by he admission conrol procedure a ime Assignmen variable - Deermines wheher radio a A of user u U is on or off a ime Rae (bis/s) allocaed o user u U by BS/AP a A a ime Maximum achievable rae (bis/s) for user u U hrough BS/AP a A a ime Normalized rae [0, 1] allocaed o user u U by BS/AP a A a ime Toal rae allocaed o user u U a ime Achievable sysem specral efficiency (bis/s/hz) for ime inerval [, +1] Toal specrum (Hz) used by he sysem Maximum daa (in bis) ha can be ransferred by radio a A of user u U during ime inerval [, +1] Vecor conaining minimum daa rae requiremen of each user u U o suppor real-ime raffic for ime inerval [, +1] Toal energy consumed (in Joules) by radio a A of user u U during ime inerval [, +1] Toal energy consumed (in Joules) by cue of user u U during ime inerval [, +1] Maximum number of usable radios for user u U for each ime sep and six or seven symbols in he ime domain o form a minimal resource allocaion uni). The minimal resource allocaion uni used by 3GPP based neworks (LTE, HSPA) is called a resource block. To unify erminology across all IP-CANs, his erm is chosen o represen a minimal resource allocaion uni for all RATs in he paper. γ u U = r u (1) κ r u = a A x ua r ua (2) Since we assume ha he amoun of specrum allocaed o each IP-CAN is consan, he oal specrum, κ, used by our sysem remains consan. So, o maximize he achievable sysem specral efficiency, he objecive of any nework opimizaion problem is o maximize he sum of he raes allocaed o each user subjec o oal resource usage consrains. This opimizaion problem has been well sudied as he Max-Sum Rae (MSR) opimizaion problem. The idea behind he MSR opimizaion objecive is o assign each resource block o he user ha can make he bes use of i. The drawback of he MSR opimizaion objecive is ha i is likely ha a few users close o he BS, and hence having excellen channels, will be allocaed all he sysem resources. As a resul, he MSR opimizaion objecive canno be used as he only objecive in any resource allocaion problem. Fairness of resource disribuion also has o be aken ino accoun.

5 AMIN e al.: BALANCING SPECTRAL EFFICIENCY, ENERGY CONSUMPTION, AND FAIRNESS 5 However, since he MSR opimizaion objecive resuls in he highes achievable sysem specral efficiency, i can be used as an upper bound in compuing he specral efficiency uiliy funcion. Le γ max represen he achievable sysem specral efficiency for ime inerval [, +1] obained by solving he MSR opimizaion problem. Similarly, assuming each available resource block is allocaed o some user, he minimum achievable sysem specral efficiency resuls when each resource block is assigned o he user wih wors conneciviy condiions. Le γ min represen his minimum achievable sysem specral efficiency for ime inerval [, +1]. Then, γ max can be used as a lower bound in compuing he specral efficiency uiliy funcion. The normalized sysem uiliy γ uil is hen compued using (3). If he achievable sysem specral efficiency equals γ max, he specral efficiency uiliy funcion corresponds o a value of 1, and if he achievable sysem specral efficiency equals γ min, he specral efficiency uiliy funcion corresponds o a value of 0. γuil = γ γmin γmax γmin 2) Fairness: The fairness meric relaes o he difference in daa raes allocaed o each user. The fairness meric can be compued for each scheduling ime sep o ensure insananeous fairness or i can be compued over long ime-scales o ensure long-erm fairness. Since we sudy heerogeneous wireless sysems ha suppor boh real-ime and bes-effor raffic, we analyze our algorihm s performance for boh insananeous fairness ha deals wih minimum daa rae suppor for real-ime raffic and long-erm fairness ha is relaed o bes-effor raffic. The firs sep in our resource allocaion procedure is an ieraive admission conrol procedure ha saisfies he minimum daa rae requiremen, T = [T1...T U ], per scheduling inerval of as many users as possible o suppor real-ime raffic. For each GRC scheduling inerval (say, 1 second), any user u U ha is allocaed resources achieves a daa rae of a leas Tu bis/s and can saisfy he needs of his/her realime applicaions. The users blocked by he admission conrol procedure are he dissaisfied users ha canno suppor he needs of heir real-ime applicaions. For each ime sep, he proporion of saisfied users is used o compue he insananeous fairness uiliy funcion, denoed θuil, as described by (4). If no users are blocked, he insananeous fairness uiliy equals 1 and if all users are blocked, he insananeous fairness uiliy equals 0. The insananeous fairness uiliy helps address he needs of real-ime raffic such as VoIP and video sreaming. For example, he G.729 codec used by VoIP raffic needs o periodically send ou variable-sized packe burss approximaely every 20 ms. Since GRC operaes on a 1 second inerval, he GRC uses average daa rae requiremens of hese variable-sized packe burss over a period of 1 second o compue he minimum daa rae requiremen per scheduling inerval. θ uil = 1 BU U (3) (4) TABLE II ENERGY CONSUMPTION NUMBERS FOR CURRENT ACCESS TECHNOLOGIES (IN JOULES) g e LTE HSDPA EVDO P,a(x KB) 0.007(x) 0.018(x) 0.018(x) 0.025(x) 0.025(x) P rec,a P assoc,a P o,a The second sep in he resource allocaion process accouns for long-erm fairness by using mechanisms from a Proporional Fair scheduler. Le φ uil represen he long-erm fairness uiliy funcion. Then we apply a direc mapping of Jain s Fairness Index [41] o he long-erm fairness uiliy funcion as presened in (5). In general, since bes-effor raffic such as FTP has very lenien or no delay consrains and because he insananeous fairness uiliy already compues he fairness meric for small ime-scales (or each scheduling sep), he long-erm fairness uiliy is compued using raes allocaed o each user for all ime seps under consideraion for any sudy (housands of seconds). As in he previous wo uiliy componens, he long-erm fairness uiliy is normalized ino he inerval [0,1]. ( u U r u) 2 φ uil = U u U ( r u) 2 (5) 3) Baery Lifeime: The final meric in our muliplearibue resource allocaion algorihm is baery lifeime. We use wo componens in modeling our energy consumpion, which is similar o a linear energy consumpion model proposed in [42, 43]. The energy consumpion for user u U during ime inerval [, +1], denoed as ω u, is compued using (6). The firs energy consumpion componen, P,a (x), relaes o he ransfer energy componen described in [43] and i depends on η ua, he maximum number of daa byes ha can be ransferred by radio a A of user u U during ime inerval [, +1]. The second energy componen, P o,a, represens he overhead energy and has wo sub-componens. The firs subcomponen, P rec,a, represens he exra energy ha is spen by RRs in reconfiguring he hardware o suppor a paricular IP- CAN. We assume an FPGA plaform as our RR plaform [44] and use he reconfiguraion energy numbers ha we derived in our previous work [23]. The second sub-componen, P assoc,a, represens he exra energy ha is spen associaing wih a new IP-CAN and is similar o he ramp energy concep used in [43]. We provide he energy consumpion numbers used in his sudy in Table II. ω uil = [P,a (ηua)+(x ua x 1 ua ) (1 x 1 ua ) P o,a ] (6) a A The goal of he baery lifeime opimizaion crierion is o prolong he ime each user remains acive in he sysem. To his end, we wan o ensure ha for each scheduling inerval [, +1], he overall energy consumed by each user in he sysem is minimized. We use he same maximum

6 6 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. XX, MAY 2013 and minimum achievable sysem specral efficiency conceps adoped in he specral efficiency uiliy funcion in compuing he baery lifeime uiliy funcion. Le ωmax represen he maximum achievable overall energy consumpion and le ωmin represen he minimum achievable overall energy consumpion for ime inerval [, +1]. Then he baery lifeime uiliy funcion for ime, denoed as ωuil, is compued using (7). If he achievable overall energy consumpion equals ωmin, he baery lifeime uiliy funcion is 1 and if he achievable overall energy consumpion equals ωmax, he baery lifeime uiliy funcion is 0. Alernaively, a proporional energy consumpion objecive could have been used as he baery lifeime opimizaion crierion in he problem formulaion, allowing users wih higher baery levels o consume more energy han users wih lower baery levels. This approach would lead o approximaely equal baery levels for each user over a long ime period. However, we do no consider his approach in his sudy and idenify i as a par of fuure work. ωuil u U = 1 ω u ωmin ωmax ωmin (7) C. Resource Allocaion Procedure In his secion, we presen he resource allocaion procedure ha is used by he GRC o come up wih cue-ip-can mappings and he rae assignmen per mapping. Since our heerogeneous wireless sysem suppors boh real-ime and bes effor raffic, he resource allocaion problem follows a wo-sep approach. In he firs sep, an ieraive admission conrol policy is implemened o saisfy minimum daa rae requiremens (for real-ime raffic) of as many users in he sysem as possible. In he second sep, he weighed specral efficiency, long-erm fairness, and baery lifeime uiliy funcions (relaed o bes-effor raffic) are maximized, subjec o minimum daa rae requiremens. Algorihm 1 describes he complee resource allocaion procedure ha is used during each ime sep. Each sep (Sep 1 and 2) in he algorihm uses a mixed ineger linear program (MILP) presened by (9) and (11) respecively. The objecive of boh MILPs is o deermine x ua (he assignmen variable) and r ua (he raes allocaed o each radio of each user). The specral efficiency, long-erm fairness and baery lifeime uiliy funcions are hen compued using hese x ua and r ua variables using (3), (5) and (7) respecively. Noe ha he baery lifeime uiliy funcion presened in (7) depends on (6) which uses an addiional variable η ua, he maximum amoun of daa (in bis) ha can be ransferred by radio a A of user u U during he scheduling inerval. Since he GRC scheduler operaes on a 1 second basis, η ua equals r ua in our sudy. The goal of he admission conrol procedure, described by Sep 1 in he algorihm, is o deermine when a user is blocked and maximize he insananeous fairness uiliy meric presened in (4) by minimizing he number of blocked users. The admission conrol procedure firs iniializes he lis of blocked users a ime (BU ) o null and ses z, he variable ha deermines he feasibiliy of saisfying realime raffic demands of each user, o be infeasible. Nex, i Algorihm 1 Muli-Aribue Resource Allocaion Sep 0: Iniializaion 1: if == 1 2: A u 1 u U 3: x 1 ua 0 u U, a A 4: ρ : end if Sep 1: Admission Conrol 6: BU, z infeasible 7: while z is infeasible 8: selec any u U, u / BU z solve P using (9) 9: if z is infeasible 10: r u,max = a A r ua,max /T u u U, u / BU 11: u drop u arg min{r u,max} 12: BU BU {u drop } 13: end if 14: end while Sep 2: Muliple-Aribue Opimizaion 15: solve MA using (11) 16: A +1 u = (1 ρ)a u + ρru recursively solves opimizaion problem P, using (9), in an effor o find a feasible soluion ha ries o saisfy he realime raffic demand of each user using consrain (9b). Noe ha in formulaing P, rua,norm is used raher han rua in consrains (9c)-(9f) o avoid non-linear problem formulaions. The relaionship beween rua and rua,norm is described by (8). This relaion removes he dependence of ru on wo variables, x ua and rua as presened in (2). Now, ru only depends on rua, as presened by (9a), as consrain (9d) makes sure ha rua,norm (and consequenly rua) is greaer han zero only if x ua equals one. Afer solving one ieraion of P, he admission conrol procedure checks wheher a feasible soluion is produced. If he soluion o P is infeasible, he user wih he wors achievable daa rae o demand raio is dropped and his user is added o he lis of blocked users (BU ) ha are assigned no resource blocks (or are assigned rae 0 as described by consrain (9e). The admission conrol procedure keeps solving P and dropping he user wih wors achievable daa rae o demand raio unil all users ha are o be allocaed resources (u U, u / BU ) can achieve a daa rae of a leas Tu bis/s. This mechanism enables he admission conrol procedure o block as few users as possible. Once a feasible soluion is produced for P, he resource allocaion procedure moves o Sep 2 of he algorihm. r ua,norm = r ua r ua,max P : max r u = a A r ua (9a) s.. ru Tu u U, u / BU (9b) rua,norm 1 a A (9c) u U (8) r ua,norm x ua u U, u / BU, a A (9d) r ua,norm = 0 u BU, a A (9e)

7 AMIN e al.: BALANCING SPECTRAL EFFICIENCY, ENERGY CONSUMPTION, AND FAIRNESS 7 rua,norm 0 u U, u / BU, a A (9f) x ua m u u U, u / BU (9g) a A x ua {0, 1} u U, u / BU, a A (9h) The final sep (Sep 2) in he algorihm comes up wih cue-ip-can mappings and he rae assignmen per mapping based on an opimizaion funcion, MA, described by (11), ha opimizes he weighed specral efficiency, long-erm fairness and energy consumpion uiliy funcions subjec o he minimum daa rae requiremens confirmed by he admission conrol procedure. The uiliy funcions described in (3) and (7) are used in MA o maximize sysem specral efficiency and minimize overall energy consumpion, respecively. For long-erm fairness, he uiliy funcion described by Jain s fairness index in (5) is non-linear and hard o solve for a largescale heerogeneous wireless sysem. As a resul, an alernaive formulaion ha uses he raio of insananeous o average daa rae described in (10) is used o maximize long-erm fairness uiliy 2. I has been shown ha allowing he user wih maximum achievable raio of insananeous o average daa rae o ransmi during each ime sep resuls in maximizing fairness over long ime scales [45]. Again, he maximum and minimum achievable raios of insananeous o average daa rae are used in (10) o scale he long-erm fairness uiliy beween 0 and 1. The algorihm iniializes he average daa rae of each user u U, denoed as A u, o 1 during he firs ime sep as described in he iniializaion sep in Algorihm 1. Afer solving he MA opimizaion problem, he algorihm updaes he average daa rae of each user over a ime window ha is dicaed by he scalar ρ. The value of his window is commonly se beween 0.05 and 0.10 [46]. We se ρ = 0.10 in our work as noed in he iniializaion sep in Algorihm 1. φ uil = [ r u A u u U [( ) r u u U A u max ( r u A u ) min ( r u A u ) ] min ] (10) MA : max (α γuil) + (β φ uil) + (τ ωuil) (11a) s.. ru Tu u U, u / BU (11b) rua,norm 1 a A (11c) u U rua,norm x ua u U, u / BU, a A (11d) rua,norm = 0 u BU, a A (11e) rua,norm 0 u U, u / BU, a A (11f) x ua m u u U, u / BU (11g) a A x ua {0, 1} u U, u / BU, a A (11h) Noe ha as saed earlier, we assume a cue can use muliple radios concurrenly. The maximum number of radios ha a cue can concurrenly use is limied by he m u variable presened in (9g) and (11g). In our problem formulaion, we 2 Noe ha φ uil presened in (10) is only used in solving MA. φ uil represening Jain s fairness index in (5) is sill used in compuing long-erm fairness uiliy. Exper 1 TABLE III AHP MATRICES DERIVED FROM EXPERT INTERVIEWS Baery Life (BL) Long-Term Fairness (LTF) Specral Efficiency (SE) Weighs Baery Life Long-Term Fairness Specral Efficiency Exper 2 Baery Life Long-Term Fairness Specral Efficiency Weighs Baery Life Long-Term Fairness Specral Efficiency assume m u o be he same for each user. There migh be cases where he value of m u can vary for differen users. For example, if a user does no have enough power o suppor more han one physical link (i.e. he user is operaing a a low baery level), hen a policy-based addiion can be included in he algorihm ha limis such a user o use only one of is radios. These policy-based decisions represen a possible exension o our curren model. The scalars α, β and τ provide he relaive imporance of each opimizaion aribue in MA and ac as conrol knobs ha allow nework operaors o achieve he desired performance objecives. The values for hese scalars are obained hrough AHP [34]. AHP is a decision analysis echnique o deermine weighs of differen uiliy aribues from decision sakeholders hrough pairwise comparisons and raings. Using AHP, we inerviewed wo expers from he cellular indusry o perform pairwise comparisons beween our uiliy aribues 3. Afer deermining which aribue is more imporan, he more imporan aribue receives a score from 1-9, wih 1 indicaing ha he wo aribues are equally imporan. These pairwise comparisons are placed in marix A, wih a ji = 1/a ij, where each row and column represens a specific aribue. Using he following equaion: Aw = λ max w, and solving for λ max, he principal eigenvalue of A, and w, he principal righ eigenvecor of A, we can normalize he enries of w by dividing by heir sum and recover he weighed values for our uiliy funcion. We asked each exper o compare he relaive imporance of baery life, fairness, and efficiency [49]. The resuls of he inerview are placed in a comparison marix, from which he principal eigenvecor is calculaed. The resuls from his calculaion and resuling weigh values are shown in Table III. From Table III, we noe ha resuls of AHP show ha boh expers had relaively similar weigh preferences. Consequenly, we use resuls derived from Exper 1 s responses in he remainder of he paper. Since he weigh used for each aribue in his paper is based on responses from only wo expers, we performed 3 While in his paper we only examine wo viewpoins, we also noe ha group decision-making and viewpoin aggregaion has been sudied in [47, 48].

8 8 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. XX, MAY 2013 TABLE IV WEIGHT SENSITIVITY ANALYSIS Carrier 1: EVDO Carrier 2: HSPA Crierion k Crieria Pair BL LTF SE BL-LTF BL-SE LTF-SE Original Weighs Carrier 1: WiFi Carrier 2: WiFi Carrier 1: WiMAX Carrier 2: WiFi Carrier 1: WiFi Carrier 2: LTE Carrier 1: WiFi Carrier 2: WiFi Crieria Pair TABLE V CRITERIA PREFERENCE VALUES Preference Range BL-LTF [0.227, 0.310] BL-SE [0.227, 0.369] LTF-SE [0.227, 0.310] BL LTF SE [0.053, 0.095] [0.053, 0.095] [0.053, 0.095] [0.608, 0.711] [0.552, 0.711] [0.608, 0.711] sensiiviy analysis o deermine how much change in weighs is required o change he relaive criicaliy of our hree opimizaion crieria (BL, LTF, SE). The criicaliy degree of a crierion is he smalles percen amoun by which he curren value of weighs mus change, such ha he exising preference ranking of he crieria will change [50]. The sensiiviy coefficien of a crierion is he reciprocal of is criicaliy degree. Using he mehodology in [50] and Exper 1 s daa presened in Table III, he preference, P, of each crierion is obained as follows: P [BL] = , P [LT F ] = , P [SE] = So he preference ranking follows he order: P [SE] > P [BL] > P [LT F ]. The minimum required change in weigh of he kh crierion o reverse he criicaliy (preference order) of wo crieria A i A j is defined as δ k,i,j such ha δ k,i,j = (P j P i )/(a jk a ik ). The resuls of his calculaion are presened in Table IV. In order o sae ha a cerain crierion is sensiive o weigh change, he change in he weigh of he k h crierion, w k, should saisfy δ k,i,j w k. However, he resuls obained in Table IV show ha none of δ k,i,j is smaller han he corresponding weighs (presened in he las row of Table IV), which means ha he hree crieria are robus and no sensiive o weigh change. In oher words, any change in he weighs will mainain he criicaliy (or preference order) of each crieria pair. This claim has been verified by changing w k, for each crierion k, in he range [w k δ k,i,j, w k + δ k,i,j ] for each A i A j crieria pair and he corresponding resuls are presened in Table V. As seen from Table V, here is no overlapping beween he preference values of he differen crieria, and so he criicaliy is preserved. This resul is as expeced inuiively as each nework exper clearly disinguishes he relaive imporance of each crierion, as presened in Table III. Fig. 4. Coverage of he simulaion opology D. Simulaion Descripion We developed a MATLAB-based simulaion model wih sufficien fideliy o demonsrae he properies of our muliaribue resource allocaion algorihm for a heerogeneous wireless sysem. We consider he presence of wo major cellular carriers ha operae muliple IP-CANs in a 2 x 2 km 2 area. Carrier 1 operaes EVDO (3G), WiMAX (4G), and IEEE g (Wi-Fi) IP-CANs and Carrier 2 operaes HSPA (3G), LTE (4G) and IEEE g (Wi-Fi) IP-CANs in our experimens. For he cellular based IP-CANs (EVDO, HSPA, WiMAX, LTE), we assume ha a single base saion serves all users in he simulaion opology. These base saions are locaed near he cener of he 2 x 2 km 2 grid. The EVDO and HSPA base saions have a coverage radius of 1.50 km and he WiMAX and LTE base saions have a coverage radius of 1.0 km. The IEEE g APs are spread hroughou he opology and have a coverage radius of 0.15 km each. There are g APs in he opology, wih 3 belonging o each carrier. We assume ha he IEEE g APs ha he cellular carriers deploy are specialized o paricipae in he cenralized global resource allocaion process by using cenralized conroller mechanisms proposed in he IEEE e MAC. To accoun for he overhead associaed wih he ransfer of messages beween BSs/APs and he cenralized GRC, we deduc 25% resource blocks from each IP-CAN compared o heir corresponding heoreical maximum. While we do no model he exac delay and signaling overhead for he log-on process each ime he access nework changes, our overhead esimae capures he reducion in efficiency ha would resul from frequen changes in he selecion of an access nework. The overall simulaion opology is presened in Fig. 4. For he proposed heerogeneous wireless sysem, we sudy wo differen use cases: 1) Use case 1 involves users ha can connec only o heir own carrier s cellular and Wi-Fi nework. 2) Use case 2 allows any user o make use of boh carriers cellular and Wi-Fi neworks. While here are economic and public policy obsacles surrounding use case 2, we assume hese obsacles will evenually be overcome. For boh use cases, he simulaion involves 100 nomadic users, 50 of which subscribe o Carrier 1 and he oher 50 o Carrier 2. All users are equipped wih 3 reconfigurable radios (using noaion

9 AMIN e al.: BALANCING SPECTRAL EFFICIENCY, ENERGY CONSUMPTION, AND FAIRNESS 9 presened in Table I, m u = 3) o suppor up o 3 IP-CANs for use case 1 and 5 IP-CANs for use case 2. They move in he simulaion opology using a random walk mobiliy model a a consan speed of 2 mph. Based on he locaion of a user in he opology, he user can connec o any available IP-CAN using one of he corresponding MCSs provided in our prior work [51]. The closer he user is o a BS/AP, he beer he signal recepion he user experiences a ha locaion. This ranslaes ino a beer MCS mapping for he specific IP-CAN under consideraion, and his governs he maximum achievable rae for each user via he corresponding BS/AP. The differen color shades in Fig. 4 represen an example MCS mapping for various IP-CANs, where he darker he MCS, he higher he order of MCS any user can use in a given locaion. Furhermore, since we do no use a deailed channel model in our sudies o deermine he MCS, we model he flucuaions in conneciviy condiions by randomly urning off each reconfigurable radio of each user for 5% of he simulaion ime. In our simulaion, each use case scenario is simulaed for 10,000 seconds. The GRC implemens a scheduler ha comes up wih cue-ip-can mappings and he corresponding rae assignmen per mapping every one second. The GRC scheduler follows he wo-sep approach described in Algorihm 1. In he second sep, while solving he MA opimizaion problem described by (11), he weighs used for specral efficiency uiliy (α), long-erm fairness uiliy (β) and baery life uiliy (τ) are 0.649, and 0.279, respecively, which correspond o he weighs derived from Exper 1 s inerview responses presened in Table III. IV. RESULTS AND ANALYSIS We firs presen resuls for when wireless daa neworks only suppor bes-effor raffic. For his case, here is no minimum daa rae requiremen for any user. In oher words, T u=0 for all users in he sysem. Since T u=0, he admission conrol procedure does no block any user for any scheduling ime sep and is no needed. As a resul, he insananeous fairness uiliy meric is no compued for his case. The overall uiliy funcion only depends on he specral efficiency uiliy (γ uil ), long-erm fairness uiliy (φ uil ) and energy consumpion uiliy (ω uil ), averaged over he enire simulaion run, and is calculaed using (12) where α = 0.649, β = 0.072, and τ = We provide he overall uiliy resuls wih each of he hree uiliy componens for use case 1 in Fig. 5. Since he overall uiliy is a normalized value in range [0,1], he resuls for use case 2 are very similar o he resuls for use case 1 and are hus omied in his paper. Opimizaion problems presened in (9) and (11), which are pars of he proposed algorihm, are solved using AMPL modeling language and CPLEX opimizaion solver [52, 53]. Overall uil,be = (α γ uil ) + (β φ uil ) + (τ ω uil ) (12) In addiion o he uiliy resuls for our muli-aribue resource allocaion algorihm, we provide resuls for four commonly used scheduling algorihms for wireless daa neworks: (#$" 49/F801;"@GF*/+FB"E:;*8B" H-+A3I/06"<1*0+/DD"E:;*8B" )*+",-./0" )123456"718/",0-9-0:-+1;"<1*0" )123)*+"<1*0" )5;:3=>0*?58/" Fig. 5. Overall Uiliy for Use Case 1, T u = 0 (i) Min Power (ii) Max-Sum Rae (iii) Proporional Fair and (iv) Max-Min Fair. Noe ha he firs hree algorihms reduce o our MA opimizaion problem if we se (i) α = 0, β = 0, τ = 1 (ii) α = 1, β = 0, τ = 0 and (iii) α = 0, β = 1, τ = 0 respecively in (11a). The Max-Min Fair resuls are obained using he progressive filling algorihm [54]. Furhermore, he Max-Sum Rae algorihm always achieves he highes sysem specral efficiency and as a resul is γ uil = 1. However, because of more conneciviy opions for use case 2, he average specral efficiency for use case 2 is 4.35 bis/s/hz compared o 3.52 bis/s/hz for use case 1. Similar o he Max-Sum Rae algorihm, he Min Power algorihm always produces he minimum possible energy consumpion and herefore is ω uil = 1. Bu he average energy consumpion per user is 9.6 mj/s for use case 1 compared o 10.4 mj/s for use case 2. All oher algorihms compue heir specral efficiency uiliy relaive o he Max-Sum Rae algorihm s specral efficiency uiliy as described by (3) and heir energy consumpion uiliy relaive o he Min Power algorihm s energy consumpion uiliy as described by (7). The overall uiliy of our muli-aribue resource allocaion algorihm is very similar o he overall uiliy of he Max- Sum Rae algorihm (0.967 compared o 0.948) as seen from Fig. 5. Since he specral efficiency uiliy is given he highes weigh in our overall uiliy funcion, his resul follows expecaions. In comparison o he Max-Sum Rae algorihm, our algorihm improves he energy consumpion uiliy (0.269 compared o 0.247) a he cos of a sligh degradaion in specral efficiency uiliy (0.648 compared o 0.649). The longerm fairness uiliy is almos he same for our algorihm and he Max-Sum Rae algorihm (approximaely 0.050). All oher algorihms (Min Power, Proporional Fair, Max-Min Fair) sacrifice specral efficiency in rying o achieve oher objecives, as seen in Fig. 5, and as a resul heir overall uiliy is much lower han he one obained by our algorihm. We now consider he case of nex-generaion heerogeneous wireless neworks ha are expeced o suppor boh realime and bes-effor raffic. In his case, he overall uiliy funcion depends on uiliy aribues ha apply o real-ime raffic and he aribues ha apply o bes-effor raffic. We equally weigh he uiliies of boh raffic ypes o compue he overall uiliy funcion. The bes-effor raffic uiliy, denoed Overall uil,be, depends on specral efficiency, long-erm fair-

10 10 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. XX, MAY 2013 ness and energy consumpion uiliies as presened in (12). The real-ime raffic depends on he insananeous fairness uiliy averaged over he enire simulaion run, denoed θ uil, and is calculaed using (4). Hence, he overall uiliy funcion is compued using (13). (#%" (#$" ("!#'"!#&"!#%" Overall uil,be+rt = 1 2 Overall uil,be θ uil (13) For he mix of real-ime and non-real-ime raffic, we presen resuls for boh use case 1 and use case 2 for varying levels of minimum daa rae requiremen of each user o suppor real-ime raffic, T u, using Fig. 6 and Fig. 7 respecively. Considering he case when T u = 512K, he overall uiliy of our algorihm for boh use cases is significanly higher han ha of any oher algorihm. For boh use cases, he overall energy consumpion uiliy and long-erm fairness uiliy of all algorihms are similar. Bu he difference in overall uiliy is obained due o insananeous fairness and specral efficiency uiliies. For use case 1, in erms of overall uiliy performance, our algorihm ouperforms he nex closes algorihm, Max-Sum Rae, by 56.7% (0.818 compared o 0.522). The specral efficiency uiliy of our algorihm for beseffor raffic decreases compared o Max-Sum Rae algorihm (0.224 compared o 0.325). Bu his happens as a resul of saisfying more real-ime raffic users. The insananeous fairness uiliy of our algorihm is significanly higher han ha of he Max-Sum Rae algorihm (0.437 compared o 0.048). For use case 2, our algorihm ouperforms he nex closes algorihm, Max-Min Fair, in erms of overall uiliy by 24.0% (0.975 compared o 0.786). The insananeous fairness uiliy of boh algorihms is 0.5. Bu he specral efficiency uiliy of our algorihm is significanly higher compared o he Max-Min Fair algorihm s specral efficiency uiliy (0.310 compared o 0.115). The resul suggess ha for fuure heerogeneous wireless sysems ha mus suppor boh real-ime and beseffor raffic, our algorihm obains he bes of boh worlds by applying he righ rade-offs in erms of achieved specral efficiency and insananeous fairness. Also noe ha for all differen levels of T u for boh use cases, our algorihm ouperforms any oher algorihm. None of he oher algorihms is suied o suppor boh bes-effor and real-ime raffic. While Max-Sum Rae and Proporional Fair algorihms are well suied for achieving good specral efficiency for bes-effor raffic, hey do no provide accepable levels of insananeous fairness. On he oher hand, he Max- Min Fair algorihm provides good insananeous fairness, bu is specral efficiency suffers significanly. Our algorihm achieves a balance in boh insananeous fairness and specral efficiency uiliies. Apar from his, here are wo addiional observaions of ineres in Fig. 6 and Fig. 7. Firs, while mos radiional algorihms provide consan overall uiliy levels and hen possibly experience sudden drops in performance (for example, Max-Min Fair algorihm for use case 2), our algorihm degrades gradually as he available resources canno saisfy he demands. Second, since use case 2 represens more conneciviy opions for each user, he resuling overall uiliy of our algorihm is considerably higher (by up o 39.4%)!#$"!" )*+",-./01"23$4&5" )*+",-./01234($5" )*+",-./01"236&'5" )*+",-./01"23(!$%5" )789:;<"=7>/1"23$4&5" )789:;<"=7>/1"234($5" )789:;<"=7>/1"236&'5" )789:;<"=7>/1"23(!$%5",0-?-0@-+7A"B7*01"23$4&5",0-?-0@-+7A"B7*01"234($5",0-?-0@-+7A"B7*01"236&'5",0-?-0@-+7A"B7*01"23(!$%5" )789)*+"B7*01"23$4&5" )789)*+"B7*01"234($5" )789)*+"B7*01"236&'5" )789)*+"B7*01"23(!$%5" Fig. 6. Overall Uiliy for Use Case 1, Variable T u (#%" (#$" ("!#'"!#&"!#%"!#$"!" F+G>7+>7+/-;G"B7*0+/GG"H@A*>I" J+/0KI"L-+G;<?@-+"H@A*>I" :?/M>07A"JNM*/+MI"H@A*>I" O-+K92/0<"B7*0+/GG"H@A*>I" )*+",-./01"23$4&5" )*+",-./01234($5" )*+",-./01"236&'5" )*+",-./01"23(!$%5" )789:;<"=7>/1"23$4&5" )789:;<"=7>/1"234($5" )789:;<"=7>/1"236&'5" )789:;<"=7>/1"23(!$%5",0-?-0@-+7A"B7*01"23$4&5",0-?-0@-+7A"B7*01"234($5",0-?-0@-+7A"B7*01"236&'5",0-?-0@-+7A"B7*01"23(!$%5" Fig. 7. Overall uiliy for Use Case 2, Variable T u )789)*+"B7*01"23$4&5" )789)*+"B7*01"234($5" )789)*+"B7*01"236&'5" )789)*+"B7*01"23(!$%5" );A@9CD0*E;>/1"23$4&5" );A@9CD0*E;>/1"234($5" );A@9CD0*E;>/1"236&'5" );A@9CD0*E;>/1"23(!$%5" );A@9CD0*E;>/1"23$4&5" );A@9CD0*E;>/1"234($5" );A@9CD0*E;>/1"236&'5" );A@9CD0*E;>/1"23(!$%5" compared o use case 1 for higher levels of T u (T u 512 kbps). So increasing he number of conneciviy opions (possibly hrough peering agreemens among several nework service providers) has significan performance benefis. V. CONCLUSIONS AND FUTURE WORK We presened an approach o managing resources in a heerogeneous wireless nework based on he 3GPP IMS archiecure ha suppors reconfigurable devices. We analyzed our muli-aribue scheduling algorihm implemened by a cenralized GRC ha considered he nework-efficiency measures of sysem specral efficiency, boh insananeous and long-erm fairness in erms of daa rae allocaed o each user in he sysem, and baery lifeime of each user in he sysem. Through Malab/CPLEX based simulaions, we showed an increase in overall uiliy of up o 57% for our algorihm compared o he nex bes algorihm. By following a wosep resource allocaion procedure, depending on he siuaion, our algorihm improves he overall sysem performance by achieving he righ rade-offs in erms of sysem specral efficiency and energy consumpion (for bes-effor raffic) or by achieving he bes rade-offs in erms of sysem specral efficiency and insananeous fairness (for real-ime raffic). As a par of fuure work, we inend o look a various incenive and economic models ha foser nework provider cooperaion o achieve an increase in overall sysem performance using he same amoun of resources ha are used oday. Also noe ha he P rec,a numbers used in his sudy were based on he assumpion ha he reconfigurable radio is compleely manufacured using he Field Programmable

11 AMIN e al.: BALANCING SPECTRAL EFFICIENCY, ENERGY CONSUMPTION, AND FAIRNESS 11 Gae Array (FPGA) echnology. As described in our previous works [23, 55], hese numbers are no absolue and can vary, perhaps due o hardware echnology advancemens or differen Applicaion Specific Inegraed Circui (ASIC) vs. FPGA implemenaion percenages for he reconfigurable radio. A scalar λ [0, 1], which we ermed as impac of reconfiguraion in our previous works, can be muliplied o P rec,a o capure he effecs of his variaion. Evaluaing differences in energy consumpion as a resul of he inroducion of impac of reconfiguraion scalar in our model also remains a par of our fuure work. In addiion, we are exending he work described in his paper by considering more sophisicaed mobiliy and channel models. REFERENCES [1] FCC, ET Docke No , Noice of Proposed Rule Making and Order, Dec [2] R. Research, HSPA o LTE-Advanced, Sep [3] I. Mansfield, Smarphone baery life has become a significan drain on cusomer saisfacion and loyaly, April [Online]. Available: hp:// [4] F. Haider, C.-X. Wang, X. Hong, H. Mischak, D. Yuan, E. Hepsaydir, and H. Haas, Specral-energy efficiency radeoff in cogniive radio neworks wih peak inerference power consrains, in Proceedings of IEEE ICCT, Sep. 2011, pp [5] C.-X. Wang, X. Hong, H.-H. Chen, and J. S. Thompson, On capaciy of cogniive radio neworks wih average inerference power consrains, IEEE Transacions on Wireless Communicaions, vol. 8, no. 4, [6] X. Kang, Y.-C. Liang, and A. Nallanahan, Opimal power allocaion for fading channels in cogniive radio neworks under ransmi and inerference power consrains, in Proceedings of IEEE ICC, [7] D. López-Pérez, I. Güvenç, G. de la Roche, M. Kounouris, T. Q. S. Quek, and J. Zhang, Enhanced inercell inerference coordinaion challenges in heerogeneous neworks, IEEE Wireless Communicaions Magazine, vol. 18, no. 3, pp , June [8] M. Sawahashi, Y. Kishiyama, A. Morimoo, D. Nishikawa, and M. Tanno, Coordinaed mulipoin ransmission/recepion echniques for LTE-advanced, Wireless Communicaions, vol. 17, no. 3, June [9] J. Deaon, M. Benonis, L. DaSilva, and R. Irwin, Supporing Dynamic Specrum Access in heerogeneous LTE+ neworks, IEEE Symposium on New Froniers in Dynamic Specrum Access Neworks (DySPAN), [10] Y. J. Zhang and K. B. Leaief, Muliuser adapive subcarrier-andbi allocaion wih adapive cell selecion for OFDM sysems, IEEE Transacions on Wireless Communicaions, vol. 3, no. 5, Sep [11] W. Rhee and J. M. Cioffi, Increase in capaciy of muliuser OFDM sysem using dynamic subchannel allocaion, in Proceedings of IEEE Vehicular Technology Conference, May 2000, pp [12] D. Tse, Muliuser diversiy in wireless neworks, Wireless Communicaions Seminar, Sanford Universiy, April [13] J.-M. Liang, J.-J. Chen, C.-W. Liu, Y.-C. Tseng, and B.-S. P. Lin, On ile-and-energy allocaion in OFDMA broadband wireless neworks, IEEE Communicaions Leers, vol. 15, no. 12, Dec [14] F. Meshkai, H. V. Poor, S. C. Schwarz, and R. V. Balan, Energyefficien resource allocaion in wireless neworks wih qualiy-of-service consrains, IEEE Transacions on Communicaions, vol. 57, no. 11, pp , Nov [15] P. Serrano, M. Hollick, and A. Banchs, On he rade-off beween hroughpu maximizaion and energy consumpion minimizaion in IEEE WLANs, Journal of Communicaions and Neworks, vol. 12, no. 2, pp , April [16] T. Bu, L. Li, and R. Ramjee, Generalized proporional fair scheduling in hird generaion wireless daa neworks, in Proceedings of IEEE INFOCOM, [17] L. Gavrilovska and V. Aanasovski, Resource managemen in wireless heerogeneous neworks (WHNs), Telsiks, Oc [18] IEEE, Sd : IEEE Sandard for Local and Meropolian Area Neworks, Par 21: Media Independen Handover Services, Jan [19] S. Buljore, H. Harada, S. Filin, P. Houze, K. Tsagkaris, O. Holland, K. Nole, T. Farnham, and V. Ivanov, Archiecure and enablers for opimized radio resource usage in heerogeneous wireless access neworks: he IEEE working group, IEEE Communicaions Magazine, vol. 47, no. 1, Jan [20] A. de la Oliva, C. J. Bernardos, M. Calderón, T. Melia, and J. C. Zúñiga, Ip flow mobiliy: smar raffic offload for fuure wireless neworks, IEEE Communicaions Magazine, vol. 49, no. 10, Oc [21] C. Bernardos, Proxy mobile IPv6 exensions o suppor flow mobiliy, IETF, draf-ief-neex-pmipv6-flow-mob-01, work in progress, Sep [22] 3rd Generaion Parnership Projec, 3GPP TS , Archiecure enhancemens for non-3gpp accesses, [23] J. Marin, R. Amin, A. Elawil, and A. Hussien, Using reconfigurable devices o maximize specral efficiency in fuure heerogeneous wireless sysems, in Proceedings of IEEE ICCCN, Aug [24] R. Marler and J. Arora, Survey of muli-objecive opimizaion mehods for engineering, Srucural and Mulidisciplinary Opimizaion, vol. 26, pp , [25] F. Kelly, A. Maulloo, and D. Tan, Rae conrol for communicaion neworks: shadow prices, proporional fairness and sabiliy, Operaional Research Sociey, vol. 49, no. 3, [26] S. Lee, K. Sriram, K. Kim, Y. H. Kim, and N. Golmie, Verical handoff decision algorihms for providing opimized performance in heerogeneous wireless neworks, IEEE Transacions on Vehicular Technology, vol. 58, no. 2, pp , [27] P. Kosmides, A. Rouskas, and M. Anagnosou, Nework selecion in heerogeneous wireless environmens, in Proceedings of IEEE Inernaional Conference on Telecommunicaions, [28] A. Calvagna and G. Di Modica, A user-cenric analysis of verical handovers, in Proceedings of he 2nd ACM inernaional workshop on Wireless mobile applicaions and services on WLAN hospos, [29] E. Sevens-Navarro and V. W. S. Wong, Comparison beween verical handoff decision algorihms for heerogeneous wireless neworks, in Proceedings of IEEE Vehicular Technology Conference, [30] P. M. Chan, R. E. Sheriff, Y. F. Hu, P. Conforo, and C. Tocci, Mobiliy managemen incorporaing fuzzy logic for a heerogeneous IP environmen, IEEE Communicaions Magazine, vol. 39, no. 12, pp , Dec [31] K. Radhika and A. V. Reddy, Nework selecion in heerogeneous wireless neworks based on fuzzy muliple crieria decision making, Inernaional Journal of Compuer Applicaions, vol. 22, no. 1, May [32] W. Luo and E. Bodanese, Opimising radio access in a heerogeneous wireless nework environmen, in Proceedings of IEEE ICC, [33] B. Bakmaz, Z. Bojkovic, and M. Bakmaz, Nework selecion algorihm for heerogeneous wireless environmen, in Proceedings of PIMRC, [34] T. Saay, How o make a decision: he analyic hierarchy process, European Journal of Operaional Research, vol. 48, [35] C. J. Chang, T. L. Tsai, and Y. H. Chen, Uiliy and game-heory based nework selecion scheme in heerogeneous wireless neworks, in Proceedings of IEEE WCNC, [36] J. Anoniou and A. Pisillides, 4G converged environmen: modeling nework selecion as a game, in Proceedings of 16h IST Mobile and Wireless Communicaions Summi, [37] D. Niyao and E. Hossain, Dynamics of nework selecion in heerogeneous wireless neworks, IEEE Transacions on Vehicular Technology, vol. 58, no. 4, pp , May [38] E. Alman, T. Boulogne, R. El-Azouzi, T. Jimenez, and L. Wyner, A survey on neworking games in elecommunicaions, Compuers and operaions Research, vol. 33, pp , [39] J. Nelson, Verizon wireless fosers global LTE ecosysems, Verizon Press Release, Feb [40] 3rd Generaion Parnership Projec, 3GPP TS V11.3.0, IP Mulimedia Subsysem (IMS); Sage 2 (Release 11), Dec [41] R. Jain, D. Chiu, and W. Hawe, A quaniaive measure of fairness and discriminaion for resource allocaion in shared compuer sysems, DEC Research Repor TR-301, Sep [42] L. M. Feeney and M. Nilsson, Invesigaing he energy consumpion of wireless nework inerface in an ad hoc neworking environmen, in Proceedings of IEEE INFOCOM, [43] N. Balasubramanian, A. Balasubramanian, and A. Venkaaramani, Energy consumpion in mobile phones: a measuremen sudy and implicaions for nework applicaions, in Proceedings of he 9h ACM SIGCOMM Inerne Measuremen Conference, [44] A. Saha and A. Sinha, An FPGA based archiecure of a novel reconfigurable radio processor for sofware defined radio, in Proceedings of Inernaional Conference on Educaion Technology and Compuer, 2009.

12 12 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. XX, NO. XX, MAY 2013 [45] P. Viswanah, D. Tse, and R. Laroia, Opporunisic beamforming using dumb anennas, IEEE Transacions on Informaion Theory, vol. 48, June [46] J. Andrews, A. Ghosh, and R. Muhamed, Fundamenals of WiMAX: Undersanding Broadband Wireless Neworking. Prenice Hall, NJ., [47] E. Forman and K. Peniwai, Aggregaing individual judgmens and prioriies wih he analyic hierarchy process, Elsevier European Journal of Operaional Research, vol. 108, no. 1, [48] R. Ramanahan and L. Ganesh, Group preference aggregaion mehods employed in AHP: An evaluaion and an inrinsic process for deriving members weighages, Elsevier European Journal of Operaional Research, vol. 79, no. 2, [49] J. D. Deaon, C. Wernz, and L. A. DaSilva, Decision analysis for Dynamic Specrum Access rules, in Proceedings of IEEE Globecom, Dec [50] E. Trianaphyllou and A. Sanchez, A sensiiviy analysis approach for some deerminisic muli-crieria decision making mehods, Decision Sciences, vol. 28, no. 1, pp , [51] J. Marin, R. Amin, A. Elawil, and A. Hussien, Limiaions of 4G wireless sysems, in Proceedings of Virginia Tech Wireless Symposium, June [52] R. Fourer, D. M. Gay, and B. W. Kernighan, AMPL: A Mahemaical Programming Language. Brooks/Cole-Thomson Learning, [53] CPLEX, ILOG CPLEX 10.0 User s Manual and Reference Manual, ILOG, S.A., [54] T. Cover and J. Thomas, Elemens of Informaion Theory. Wiley, New York, NY, [55] R. Amin, J. Marin, A. Elawil, and A. Hussien, Specral Efficiency and Energy consumpion radeoffs for reconfigurable devices in heerogeneous wireless sysems, in Proceedings of IEEE WCNC, April Rahul Amin is a PhD Candidae in he Holcombe Deparmen of Elecrical and Compuer Engineering a Clemson Universiy, USA. Since 2009, he has been a Research Assisan in he Neworking Lab for he School of Compuing a Clemson Universiy and has focused on radio resource managemen sraegies for nex generaion heerogeneous wireless neworks. Before his PhD work, he worked as a Research Assisan in Deparmen of Elecrical and Compuer Engineering a Clemson Universiy as a Masers suden and collaboraed on a research projec wih BMW IT research cener in Greenville, SC where he focused on vehicular broadband conneciviy projec using a WiMAX nework deploymen. Before graduae school, he held an inernship posiion a Adran where he worked in he design verificaion esing and sraegic sofware groups ha focused on he developmen of Adran s swiches, rouers and muliplexer producs. Dr. Jim Marin is an Associae Professor in he School of Compuing a Clemson Universiy. His research ineress include broadband access, wireless neworks, Inerne proocols, and nework performance analysis. Curren research projecs include heerogeneous wireless sysems and DOCSIS 3.x cable access neworks. He has received funding from NSF, NASA, he Deparmen of Jusice, Cisco, IBM, CableLabs, and BMW. Dr Marin received his Ph.D. from Norh Carolina Sae Universiy. Prior o joining Clemson, Dr Marin was a consulan for Garner, and prior o ha, a sofware engineer for IBM. Juan Deaon is a Wireless Sysems Researcher a he Idaho Naional Lab (INL), a recen Ph.D. Graduae from Virginia Tech, and is currenly invesigaing echniques for specrum modeling and predicion o develop a cogniive MAC. His mos recen conribuions propose: he Specrum Accounabiliy Framework, which enables DSA in 4G wireless neworks and beyond. Resuls were published in he presigious 2011/2012 Inernaional Symposium on New Froniers in Dynamic Specrum Access Neworks. Before graduae school, Juan researched opions for wireless airborne emergency communicaions, vulnerabiliies of VoIP applicaions, and securiy issues wih indusrial wireless sysems for he INL. Before he INL, Juan worked for Moorola s CDMA nework division where he wroe sysem requiremens, esed, and coordinaed new equipmen deploymens ino commercial markes and was he recipien of he 2003 CEO award for voluneerism. His paened work includes: mobile adverising, nework upgrades, and emergency communicaion sysems. Luiz A. DaSilva is a Professor in he Bradley Deparmen of Elecrical and Compuer Engineering a Virginia Tech, USA. He also holds he Sokes Professorship in Telecommunicaions in he Deparmen of Elecronic and Elecrical Engineering a Triniy College Dublin. His research focuses on disribued and adapive resource managemen in wireless neworks, and in paricular cogniive radio neworks and he applicaion of game heory o wireless neworks. Prof. DaSilva is currenly a principal invesigaor on research projecs funded by he Naional Science Foundaion in he Unied Saes, he Science Foundaion Ireland, and he European Commission under Framework Programme 7. He is a co-principal invesigaor of CTVR, he Telecommunicaions Research Cenre in Ireland. He has co-auhored wo books on wireless communicaions and in 2006 was named a College of Engineering Faculy Fellow a Virginia Tech. Amr Hussien received his B.Sc. and M.Sc. degrees from he Elecronics and Communicaions Deparmen, Cairo Universiy, Egyp in 2006 and 2008 respecively. He is currenly a PhD Candidae a Elecrical Engineering and Compuer Science (EECS) Deparmen, Universiy of California, Irvine. He has been a Research Assisan a he Wireless Sysems and Circuis Laboraory (WSCL). He has held inernship posiions a Mindspeed, Newpor Beach, CA where he worked in he PHY implemenaion of dual mode 3G/4G projec. His curren research ineress include low power reconfigurable archiecures for fuure wireless sysems and hardware error resilien low power archiecures for wireless sysems. Ahmed Elawil is an Associae Professor a he Universiy of California, Irvine. He received he Docorae degree from he Universiy of California, Los Angeles, in Since 2005, he has been wih he Deparmen of Elecrical Engineering and Compuer Science, Universiy of California, Irvine. He is he founder and direcor of he Wireless Sysems and Circuis Laboraory (hp://newpor.eecs.uci.edu/ aelawil/). His curren research ineress are in low power digial circui and signal processing archiecures for wireless communicaion sysems. He received several disinguished awards, including he NSF CAREER award in 2010 supporing his research in low power sysems.